THE 60-SECOND VERSION
● Reddit lost 86% of its ChatGPT citations in three days. The share went to vendor docs, not other forums.
● Companies now run 13 AI agents each, up from 5. Seven in ten support sessions are autonomous.
● NVIDIA reports Q2 tonight, after market close. The results move valuations across the whole sector.
● 5 days to the 31 August double deadline.
Reddit nearly vanished from ChatGPT, and the destination is the story
Promptwatch tracking shows reddit.com held a steady 3.83 percent share of all ChatGPT Search citations from 18 July through 7 August. By 14 to 17 August that average was 0.52 percent — an 86.4 percent relative collapse. In April, Reddit had been the single most-cited domain on the platform.
The trigger was a change on 8 August that OpenAI never announced: ChatGPT Search began using the site: operator at scale, with domain-scoped queries jumping from 0.37 percent to 16.8 percent of all background searches in a single day.
Google AI Overviews fell 11 percent over the same window and AI Mode about 30 percent — gradual, not a cliff. This was ChatGPT-specific. And a separate model-version analysis found the lost share moved to vendor documentation and first-party brand pages rather than to other discussion platforms.
Full breakdown: the data, the caveats, and what publishers should check →
Enterprises now run 13 agents each
Salesforce's Agentic Enterprise Index, drawn from production activity across 400 businesses, puts agents per organisation at 13, up from five in early 2025. Build time fell 53 percent, employee sessions tripled, and seven in ten customer-service sessions are now handled autonomously.
The number worth noticing is that escalations stayed steady. Autonomous handling rose without more customers ending up with a human anyway — which is what separates real deflection from a hidden queue.
Full breakdown, and why it does not contradict the coding-agent data →
Claude chat and Cowork now share one memory
Anthropic merged the two memory systems on 25 August. Hand Cowork a task and it starts from what Claude learned in your chats, and anything from a Cowork task carries back. Memories now also form during a conversation rather than by reviewing it at the end.
Two details most coverage skips. It only works when Cowork runs in the cloud — local sessions use no memory at all. And there is no option to keep the two memories separate; Anthropic confirmed they share one system. Memory is on by default for Free, Pro and Max, off by default for Team and Enterprise.
Full breakdown: the controls, the defaults, and what to check in Settings →
OpenAI published its first chip benchmarks, one day early
At Hot Chips on 25 August, OpenAI released the first benchmark results for Jalapeño, its custom inference chip built with Broadcom. The claims: 1.5 to 1.9 times more work per watt and 1.7 to 3.6 times lower latency than NVIDIA GB200 and GB300, rising to 2.1 to 4.1 times on interactive workloads.
Four caveats travel with those numbers. OpenAI ran the tests itself, with SemiAnalysis verifying the runs in person but not executing the full suite. Jalapeño cannot train models at all. It was not compared against Vera Rubin, NVIDIA's next generation. And against a GB300 running multi-token prediction — the configuration NVIDIA deployments commonly use in production — the efficiency lead falls to roughly 1.5 times.
Full breakdown: the numbers, the caveats, and what it means for inference pricing →
NVIDIA reports tonight
NVIDIA releases Q2 results after market close today. As the largest AI company by market value, its numbers move valuations across the sector rather than just its own share price — and the datacentre segment is the line everyone reads as a proxy for how much compute is actually being bought.
Nothing to act on before the print. Worth knowing it is happening. Not investment advice.
Five days to the double deadline
| Date |
What happens |
| Aug 31 |
Claude Sonnet 5 goes $2 to $3 per million input, $10 to $15 output, plus a tokenizer change adding 10 to 35 percent tokens on code |
| Aug 31 |
GPT-5.4 and GPT-5.4 mini leave Codex for ChatGPT sign-in users. Still available via API key |
| Aug 31 |
kimi-k2.5 and moonshot-v1 sunset, migrate to kimi-k3 |
| Early-mid Sept |
Grok 4.7 window per Musk. Still no model ID at docs.x.ai |
| Oct 1 |
OpenAI vs Apple hearing |
| Oct 24 |
deepseek-chat and deepseek-reasoner deprecated |
| ~Nov 21 |
GPT-5.6 Sol promotional pricing ends, $4/$20 back to $5/$30 |
Still open
- CVE-2025-62593 in Ray. Anything below 2.52.0 remains exposed, and the attack path is a developer's browser. What to check →
- OpenAI's public S-1. Filed confidentially 8 June, still not on EDGAR.
- France's regulator estimates AI-generated summaries have cut traffic to publisher sites by 33 to 38 percent, reviving the earlier dispute over Google's use of French news content.
FAQ
Why did Reddit lose its ChatGPT citations?
On 8 August ChatGPT Search began using the site: operator at scale, scoping queries to named domains rather than searching the open web. Reddit's share fell from 3.83 percent to 0.52 percent by 14 to 17 August. OpenAI has not commented.
Did the same happen on Google?
No. Reddit fell 11 percent in AI Overviews and about 30 percent in AI Mode over the same window, both gradual. The cliff was ChatGPT-specific.
How many AI agents does a typical company run now?
Thirteen, up from five in early 2025, per Salesforce data covering production activity across 400 businesses.
What changes on 31 August?
Claude Sonnet 5 rises to $3 in and $15 out per million with a tokenizer change alongside, and GPT-5.4 and GPT-5.4 mini leave Codex for ChatGPT sign-in users. Both remain available via API.